On 16/04/21 15:51, Vincent Guittot wrote:
> Le jeudi 15 avril 2021 � 18:58:46 (+0100), Valentin Schneider a �crit :
>> +
>> +/*
>> + * What does migrating this task do to our capacity-aware scheduling
>> criterion?
>> + *
>> + * Returns 1, if the task needs more capacity than the dst CPU can provide.
>> + * Returns 0, if the task needs the extra capacity provided by the dst CPU
>> + * Returns -1, if the task isn't impacted by the migration wrt capacity.
>> + */
>> +static int migrate_degrades_capacity(struct task_struct *p, struct lb_env
>> *env)
>> +{
>> + if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
>> + return -1;
>> +
>> + if (!task_fits_capacity(p, capacity_of(env->src_cpu))) {
>> + if (cpu_capacity_greater(env->dst_cpu, env->src_cpu))
>> + return 0;
>> + else if (cpu_capacity_greater(env->src_cpu, env->dst_cpu))
>> + return 1;
>> + else
>> + return -1;
>> + }
>
> Being there means that task fits src_cpu capacity so why testing p against
> dst_cpu ?
>
Because if p fits on src_cpu, we don't want to move it to a dst_cpu on
which it *doesn't* fit.
>> +
>> + return task_fits_capacity(p, capacity_of(env->dst_cpu)) ? -1 : 1;
>> +}
>
> I prefer the below which easier to read because the same var is use
> everywhere and you can remove cpu_capacity_greater.
>
> static int migrate_degrades_capacity(struct task_struct *p, struct lb_env
> *env)
> {
> unsigned long src_capacity, dst_capacity;
>
> if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
> return -1;
>
> src_capacity = capacity_of(env->src_cpu);
> dst_capacity = capacity_of(env->dst_cpu);
>
> if (!task_fits_capacity(p, src_capacity)) {
> if (capacity_greater(dst_capacity, src_capacity))
> return 0;
> else if (capacity_greater(src_capacity, dst_capacity))
> return 1;
> else
> return -1;
> }
>
> return task_fits_capacity(p, dst_capacity) ? -1 : 1;
> }
>
I'll take it, thanks!
>
>> +
>> #ifdef CONFIG_NUMA_BALANCING
>> /*
>> * Returns 1, if task migration degrades locality
>> @@ -7672,6 +7698,15 @@ int can_migrate_task(struct task_struct *p, struct
>> lb_env *env)
>> if (tsk_cache_hot == -1)
>> tsk_cache_hot = task_hot(p, env);
>>
>> + /*
>> + * On a (sane) asymmetric CPU capacity system, the increase in compute
>> + * capacity should offset any potential performance hit caused by a
>> + * migration.
>> + */
>> + if ((env->dst_grp_type == group_has_spare) &&
>
> Shouldn't it be env->src_grp_type == group_misfit_task to only care of misfit
> task case as
> stated in $subject
>
Previously this was env->idle != CPU_NOT_IDLE, but I figured dst_grp_type
could give us a better picture. Staring at this some more, this isn't so
true when the group size goes up - there's no guarantees the dst_cpu is the
one that has spare cycles, and the other CPUs might not be able to grant
the capacity uplift dst_cpu can.
As for not using src_grp_type == group_misfit_task, this is pretty much the
same as [1]. CPU-bound (misfit) task + some other task on the same rq
implies group_overloaded classification when balancing at MC level (no SMT,
so one group per CPU).
[1]: http://lore.kernel.org/r/jhjblcuv2mo.mog...@arm.com
